The 35th annual conference on Astronomical Data Analysis Software & Systems 2025 in Görlitz, Germany

We present a novel representation learning framework for the Gaia XP DR3 dataset that leverages two advanced tools: Spherinator and HiPSter. Spherinator provides a method for learning compact representations of high-dimensional data, including images, point clouds, data cubes, time series, and spectra. Our training process uses variational autoencoders with hyperspherical latent spaces to efficiently and robustly extract physically meaningful parameterizations of data properties. Our approach explicitly incorporates uncertainties from experiments into representation learning, which produces more robust and physically consistent latent representations. HiPSter generates and serves HiPS-based (hierarchical progressive surveys) representations of learned features, enabling the scalable visualization and exploration of the latent space using Aladin-Lite.
We demonstrate the scientific potential of our method using empirical results from Gaia XP DR3 and showcase the effectiveness of cross-disciplinary tools developed under the EU SPACE initiative to enhance data-driven astronomy.